5 ms·

It makes no mention of the physical layer, but spinning disks need to 1. increase their device bandwidth SAS only does what, 1200MB/s ? Networks are at 200-400

It makes no mention of the physical layer, but spinning disks need to 1. increase their device bandwidth SAS only does what, 1200MB/s ? Networks are at 200-400G right now. And an individual SAS device can only do 12Gbits. 2. take predicates, even if the interface can handle 4-16GB/s your internal bandwidth and the ability to take a predicate means the drive can start understanding the data it stores 3. HDD physical media bandwidth needs to increase. Hard drives should move to 1-4U sizes, spin slower, have massively larger platters and multiple read heads. If we assume 200GB/cm^2, we come out to ballparkish numbers for modern drives if we assume 5 platters. Lets just say that is true. A single side of a 30cm platter is equivalent to 5 platters, two sides so 10 surfaces of a 9cm platter. Lets ignore the space taken up by the hub. So a 19inch rack sized storage device is going to have 10x the storage capacity. It might have 10x the HD heads? Is more a chicken all in a single egg and ... I think would have to do something more impressive to get economies of scale. The current 3.5" drive is going to pack a lot more than 10x to the 1u. So I think the current target I'd like to see drives go is a higher bandwidth access to the device, larger buffers, ability to send search predicates and a higher level protocol. We need HDD with NVMe. One should be able to do 4GB/s to an HDD for at least 10 seconds. All HDDs should be hybrid except for DVR drives as they only see there sustained write rate.

wtallis 5y ago
Hard drives are still first and foremost optimized for low cost per TB. Almost everything you're asking for would work against that goal, even if you could get economies of scale on these specialized drives. You would end up with little or no cost savings relative to QLC NAND, sequential read performance and random IO performance that would still be worse than QLC, and potentially higher latency than current hard drives. Seagate has been pursuing dual-actuator drives because density increases have been causing a decline in performance per TB, making the performance characteristics of big hard drives relatively more tape-like. It's a first small step in the direction of your proposals, but it doesn't seem to be a big hit so far. That indicates taking those ideas to their logical extremes is probably very uneconomical.
- sitkack 5y ago
  Thanks for your analysis, I appreciate it. I was kinda saying, "hey why don't we have LaserDisc sized hard drive platters, they would be HUGE!". Using lots of smaller disks has so many advantages after I thought about it. One being additive bandwidth, if a single disk is 0.2GB/s if I need 400GB/s I know I need at least 200 disks to hit average write rates of 0.4TB/s. Speaking of tape like behavior, it would seem that SMR would benefit a whole bunch from multiple actuators. It seems like rather than having dual actuators, as that doubles the number of simultaneous streams, it would help with lots of shared workload, but really that problem should be solved somewhere else -OR- widening the read head by some multiple. Instead of using a record player needle, use a rake. Interleave the tongues so there are like k tracks between splines. Also reduces the actuator travel distance. We should think of individual drives not as DIMMs but the individual chips themselves. They can be remixed into what ever your workload demands.
throw63738 5y ago
large plate has higher centripetal force, vibrations etc.
numpad0 5y ago
HDDs had larger platters and multiple heads, they evolutionarily lost them. I could be completely wrong but judging by audible noises disks make I suppose vibrations make it hard to actually run multiple heads.
- namibj 5y ago
  For the latter, it should suffice to use a monolithic closed loop control that intrinsically compensates for induced torque coupled between actuators through the base.
mike_d 5y ago
> Hard drives should move to 1-4U sizes Consumers will be super excited that their desktop computers now take twice as much space.
- 6nf 5y ago
  Consumers are quickly moving to all SSD
- ianlevesque 5y ago
  What consumers? Presently if you pay a lot you get an SSD, if you pay a little you get eMMC. I haven’t seen an HDD in years now.
chefkoch 5y ago
Who would need this? Why build bigger drives if you can do the same with the existing drives, RAID and SSD caching?
- katbyte 5y ago
  Because you need to store 100s of TBs?
  - chefkoch 5y ago
    And you can't this with zfs and the likes?
    - katbyte 5y ago
      the $/Tb is far less with HDDs than SSDs and that doesn't show any sign of changing soon. HDD can go as low as 0.014$ per gb, while SSDs at their lowest are 0.070$ - 4x more storage for your $ and it also doesn't expire due to writes. https://diskprices.com/?locale=us&condition=new,used&units=gb&capacity=16-&disk_types=external_hdd,external_hdd25,internal_hdd,internal_hdd25,internal_sshd,internal_sas,external_ssd,internal_ssd,m2_ssd,m2_nvme https://diskprices.com/?locale=us&condition=new,used&units=g...
sitkack 5y ago
*edit? I am replying to myself in hopes of replying to everyone who responded. You are all right and much smarter than me. I kinda outlined that lots of aggregated 3.5" (9cm) drives are better than some sorta boomer 30cm platters. I'll admit it was kinda for aesthetics.

wtallis 5y ago

Hard drives are still first and foremost optimized for low cost per TB. Almost everything you're asking for would work against that goal, even if you could get economies of scale on these specialized drives. You would end up with little or no cost savings relative to QLC NAND, sequential read performance and random IO performance that would still be worse than QLC, and potentially higher latency than current hard drives. Seagate has been pursuing dual-actuator drives because density increases have been causing a decline in performance per TB, making the performance characteristics of big hard drives relatively more tape-like. It's a first small step in the direction of your proposals, but it doesn't seem to be a big hit so far. That indicates taking those ideas to their logical extremes is probably very uneconomical.

sitkack 5y ago
Thanks for your analysis, I appreciate it. I was kinda saying, "hey why don't we have LaserDisc sized hard drive platters, they would be HUGE!". Using lots of smaller disks has so many advantages after I thought about it. One being additive bandwidth, if a single disk is 0.2GB/s if I need 400GB/s I know I need at least 200 disks to hit average write rates of 0.4TB/s. Speaking of tape like behavior, it would seem that SMR would benefit a whole bunch from multiple actuators. It seems like rather than having dual actuators, as that doubles the number of simultaneous streams, it would help with lots of shared workload, but really that problem should be solved somewhere else -OR- widening the read head by some multiple. Instead of using a record player needle, use a rake. Interleave the tongues so there are like k tracks between splines. Also reduces the actuator travel distance. We should think of individual drives not as DIMMs but the individual chips themselves. They can be remixed into what ever your workload demands.

throw63738 5y ago

large plate has higher centripetal force, vibrations etc.

numpad0 5y ago

HDDs had larger platters and multiple heads, they evolutionarily lost them. I could be completely wrong but judging by audible noises disks make I suppose vibrations make it hard to actually run multiple heads.

namibj 5y ago
For the latter, it should suffice to use a monolithic closed loop control that intrinsically compensates for induced torque coupled between actuators through the base.

mike_d 5y ago

> Hard drives should move to 1-4U sizes Consumers will be super excited that their desktop computers now take twice as much space.

6nf 5y ago
Consumers are quickly moving to all SSD
ianlevesque 5y ago
What consumers? Presently if you pay a lot you get an SSD, if you pay a little you get eMMC. I haven’t seen an HDD in years now.

chefkoch 5y ago

Who would need this? Why build bigger drives if you can do the same with the existing drives, RAID and SSD caching?

katbyte 5y ago
Because you need to store 100s of TBs?
- chefkoch 5y ago
  And you can't this with zfs and the likes?
  - katbyte 5y ago
    the $/Tb is far less with HDDs than SSDs and that doesn't show any sign of changing soon. HDD can go as low as 0.014$ per gb, while SSDs at their lowest are 0.070$ - 4x more storage for your $ and it also doesn't expire due to writes. https://diskprices.com/?locale=us&condition=new,used&units=gb&capacity=16-&disk_types=external_hdd,external_hdd25,internal_hdd,internal_hdd25,internal_sshd,internal_sas,external_ssd,internal_ssd,m2_ssd,m2_nvme https://diskprices.com/?locale=us&condition=new,used&units=g...

sitkack 5y ago

*edit? I am replying to myself in hopes of replying to everyone who responded. You are all right and much smarter than me. I kinda outlined that lots of aggregated 3.5" (9cm) drives are better than some sorta boomer 30cm platters. I'll admit it was kinda for aesthetics.